At FMS 2026 in Santa Clara this week, SK Hynix and Samsung showed two very different ways to fight the same battle, and the difference comes down to a classic standards war dynamic: when a new technology needs the rest of an industry to adopt it, having allies often matters more than having the better spec sheet.
SK Hynix’s answer to the “memory wall” — the gap between how fast AI chips can compute and how fast memory can feed them data — is a new layer called HBF (High Bandwidth Flash), stacking NAND like HBM stacks DRAM to add capacity at close to HBM-like speed. It didn’t build this alone. SK Hynix formed a consortium with Sandisk in February, and by the time the two companies released the actual technical specification at FMS, Google and AI chip startup Tenstorrent had joined too. The spec was published through the Open Compute Project as an open standard, not a proprietary format, so any chipmaker can design around it.
Samsung’s answer, zHBM, takes the opposite structural bet: instead of adding a new memory tier, it stacks HBM directly on top of the AI chip itself, cutting the data’s travel distance and promising roughly 8x the performance of HBM5. But Samsung built it alone, unveiled it as a concept model, and gave no commercialization timeline or pricing — a sharp contrast to SK Hynix’s consortium, which already has a published spec and named allies in the pipeline.
Neither approach is obviously wrong; solo strategies can move faster and keep the IP in-house, which is exactly the tradeoff Samsung is making. But history’s standards wars — VHS beating Betamax despite arguably worse picture quality — tend to reward whichever side gets more companies building around the same rulebook, not whichever side ships the flashier demo first. It’s the same dynamic behind IPO float scarcity in China’s own memory sector: capability alone rarely wins a standards fight.
My take: In a standards war, the winning spec is usually the one more companies agreed to build around.
Not advice. Just how I see it.